Data center air conditioning equipment load prediction method and system

Abstract

According to the data center air conditioning equipment load prediction method and system, through triple technical innovation of time-space thermal coupling modeling, physical constraint embedding and dynamic model prediction control, energy efficiency and stability are remarkably improved, an STGCN-LSTM hybrid model is adopted to fuse a spatial topological relation and time-varying environment disturbance, and compared with traditional LSTM prediction, the error is lower; the frequency of the compressor is dynamically adjusted by combining an MPC control strategy with load prediction, overrunning is avoided, and the energy consumption of the air conditioner is effectively reduced; the physical constraint correction layer ensures that the output accords with thermodynamic conservation, the temperature fluctuation range is narrowed to be lower, and the starting and stopping times of equipment are obviously reduced; the digital twin module realizes fault scene simulation, and the operation and maintenance response efficiency is obviously improved; the method is especially suitable for a high-density data center, and provides key technical support for green transformation of ICT infrastructure under the dual-carbon target.

Description

Technical Field

[0001] This invention relates to the field of data center cooling technology, and in particular to a method and system for predicting the load of data center air conditioning equipment. Background Technology

[0002] With the rapid development of cloud computing and artificial intelligence technologies, data centers are continuously expanding in scale, and their energy consumption now accounts for 3% of global electricity consumption. Air conditioning systems, as the core heat dissipation facilities of data centers, account for more than 40% of the total energy consumption. Existing air conditioning control systems mostly adopt feedback regulation based on fixed thresholds, relying on real-time temperature monitoring to adjust cooling output. However, the heat load of data centers is highly dynamic: server workloads fluctuate significantly due to business peaks and valleys, and outdoor weather conditions change day and night, leading to two major bottlenecks in traditional methods:

[0003] First, there is lag and insufficient stability. Feedback control only triggers cooling after the temperature exceeds the limit, but thermal inertia causes temperature fluctuations to exceed ±2℃. Frequent compressor starts and stops accelerate equipment aging; a large data center reported that this led to a 35% increase in failure rate. Second, there is significant energy waste. To cope with sudden loads, air conditioners often operate in oversaturated mode, with the actual load rate only 60-70%, resulting in annual redundant power consumption exceeding 2 million kilowatt-hours.

[0004] Existing improvement solutions fall into two categories: one is statistical models based on historical data, but these struggle to capture nonlinear thermal coupling effects; the other is pure machine learning methods that ignore the laws of thermodynamic conservation, potentially leading to predictions that violate energy balance. For example, patent document CN 118705730A fits outdoor temperature using a cosine function, but fails to consider the complex thermal interactions within the data center, resulting in a prediction error as high as 15%.

[0005] Therefore, there is an urgent need for an intelligent prediction method that integrates physical mechanisms and spatiotemporal characteristics to achieve accurate load tracking and dynamic energy consumption optimization. Summary of the Invention

[0006] The purpose of this invention is to provide a method and system for predicting the load of data center air conditioning equipment, so as to solve the problems existing in the prior art.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] This invention provides a method for predicting the load of data center air conditioning equipment, comprising the following steps:

[0009] S1. Collect historical operational data from the data center;

[0010] S2. Perform spatiotemporal alignment on the collected historical operation data, construct a spatiotemporal thermal coupling matrix, and use the isolated forest algorithm to clean up abnormal data;

[0011] S3. Establish a hybrid prediction model;

[0012] S4. Use a Seq2Seq structure to perform multi-step rolling prediction and generate prediction results;

[0013] S5. Generate air conditioning control commands based on the prediction results, and dynamically adjust the compressor frequency and valve opening.

[0014] Preferably, in step S1, the historical operating data includes server power consumption P. it Temperature difference between cabinet inlet and outlet chiller unit power P cool Outdoor temperature and humidity T out Air conditioning water supply temperature T supply Air conditioning return water temperature T return .

[0015] Preferably, in step S2, the spatiotemporal thermal coupling matrix is:

[0016]

[0017] in, Let be the heat load intensity index of cabinet (i,j) at time t. α is the server power consumption weighting coefficient, with an initial value of 0.7. Let be the total power consumption of the servers in rack (i,j) at time t, and β be the thermal gradient weighting coefficient, with an initial value of 0.3. Let t be the temperature difference between the front and back ends of the cabinet (i,j).

[0018] Its construction methods include:

[0019] S21. Define the connection relationship between adjacent cabinets based on the data center topology graph G=(V,E), where V is a cabinet node and E is an adjacent edge;

[0020] S22. The total power consumption of the servers in the rack is combined with the temperature difference between the front and back ends of the rack by weighted summation. The server power consumption weight coefficient and the thermal gradient weight coefficient are optimized by backpropagation.

[0021] Preferably, step S3 includes:

[0022] S31. Spatial features are extracted using a spatiotemporal graph convolutional layer, with the following formula:

[0023]

[0024] in, To add a self-loop adjacency matrix, Let X be a degree matrix. (l) Let W be the input feature matrix of the l-th layer. (l)Let σ be the trainable weight matrix of the l-th layer, and σ be the ReLU activation function.

[0025] S32. Extract temporal features using an adaptive forgetting gate LSTM layer, the formula is as follows:

[0026] f t =σ(W f ·[h t-1 ,x t ]+b f +C f ·ΔT env );

[0027] Among them, f t W is the output value of the forget gate at time t. f h is the forget gate weight matrix. t-1 Let x be the hidden state vector at time t-1. t Let b be the input feature at time t. f For the forget gate bias term, C f Let ΔT be the environmental disturbance sensitivity coefficient. env The outdoor temperature difference disturbance factor;

[0028] S33. Physical constraint correction layer, the formula is:

[0029]

[0030] in, The corrected load prediction value at time t. Let ∑P be the original load forecast value at time t. IT Let P be the total power consumption of IT equipment at time t, η be the COP coefficient of the air conditioning system, and P be the total power consumption of IT equipment at time t. cool_max This represents the maximum theoretical cooling capacity of the refrigeration system.

[0031] Preferably, in step S32, the outdoor temperature difference disturbance factor ΔT env The calculation method is as follows:

[0032]

[0033] Preferably, in step S33, the COP coefficient η of the air conditioning system is obtained by fitting historical data, with the following constraints:

[0034] |∑P IT -η·P cool |≤ε, where ε is the allowable error threshold for energy conservation.

[0035] Preferably, in step S4, the formula for the multi-step rolling prediction is:

[0036] [Y t+1 ,...,Y t+k] = Decoder(Encoder(H t-n:t ),C phys );

[0037] Among them, Y t+1 ,...,Y t+k Given a future k-step load prediction sequence, Decoder is a decoder network, Encoder is an encoder network, and H... t-n:t For the historical spatiotemporal thermal coupling matrix sequence, C phys This is the physical constraint vector that includes the pipeline resistance curve;

[0038] The multi-step rolling prediction uses the TeacherForcing mechanism to train the Seq2Seq model.

[0039] Preferably, in step S5, the objective function of the control pointer is:

[0040]

[0041] Among them, T k Let k be the predicted temperature, ω1 be the temperature setpoint, ω1 be the temperature stability weight (default value is 0.6), and Δf be the temperature setpoint. k ω is the compressor frequency change at time k, ω2 is the equipment action penalty weight with a default value of 0.4, and N is the optimization time domain length.

[0042] The present invention also provides a data center air conditioning equipment load prediction system, comprising:

[0043] The data acquisition module is equipped with multi-source sensors and data interfaces to collect historical operational data from the data center.

[0044] The data processing module is used to perform spatiotemporal alignment on the collected historical operational data, construct a spatiotemporal thermal coupling matrix, and use the isolated forest algorithm to clean up abnormal data.

[0045] Model building module: used to integrate the STGCN-LSTM hybrid model and the physical constraint correction layer to build a hybrid prediction model;

[0046] Prediction module: Used to perform multi-step rolling prediction using a Seq2Seq structure and generate prediction results;

[0047] Control module: Used to generate air conditioning control commands based on prediction results, and dynamically adjust compressor frequency and valve opening;

[0048] The digital twin module is used to generate 3D heat maps to render the temperature distribution in the computer room in real time, and to control the robustness of the strategy in simulation tests of load change scenarios.

[0049] Preferably, the data acquisition module includes:

[0050] Temperature sensors are deployed at cabinet entrances and exits, cold aisles, and air conditioning water pipes;

[0051] A power consumption monitoring device, which is connected to IT equipment and a chiller unit;

[0052] The Internet of Things (IoT) gateway integrates BMS system data and weather station API, with a sampling frequency of 1 minute / time for environmental data and 5 seconds / time for device power consumption.

[0053] The present invention achieves the following beneficial technical effects compared to the prior art:

[0054] This invention provides a load prediction method and system for data center air conditioning equipment. Through three innovative technologies—spatiotemporal thermal coupling modeling, physical constraint embedding, and dynamic model predictive control—it significantly improves energy efficiency and stability. The STGCN-LSTM hybrid model integrates spatial topology relationships and time-varying environmental disturbances, resulting in lower prediction errors compared to traditional LSTM. An MPC control strategy combined with load prediction dynamically adjusts the compressor frequency, avoiding over-sizing and effectively reducing air conditioning energy consumption. A physical constraint correction layer ensures that the output conforms to thermodynamic conservation, narrowing the temperature fluctuation range and significantly reducing the number of equipment start-ups and shutdowns. A digital twin module enables fault scenario simulation, significantly improving operational response efficiency. This method is particularly suitable for high-density data centers, providing key technical support for the green transformation of ICT infrastructure under the "dual carbon" target. Attached Figure Description

[0055] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0056] Figure 1 The flowchart of the data center air conditioning equipment load prediction method provided by the present invention is shown. Detailed Implementation

[0057] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0058] The purpose of this invention is to provide a method and system for predicting the load of data center air conditioning equipment, so as to solve the problems existing in the prior art.

[0059] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0060] Example 1:

[0061] This embodiment provides a method for predicting the load of data center air conditioning equipment, such as... Figure 1 As shown, it includes the following steps:

[0062] S1. Collect historical operating data from the data center; historical operating data includes server power consumption P. it Temperature difference between cabinet inlet and outlet chiller unit power P cool Outdoor temperature and humidity T out Air conditioning water supply temperature T supply Air conditioning return water temperature T return ;

[0063] S2. Perform spatiotemporal alignment on the collected historical operation data, construct a spatiotemporal thermal coupling matrix, and use the isolated forest algorithm to clean up abnormal data;

[0064] This can include: data alignment: multi-source data synchronization based on timestamps; feature engineering: spatial features, rack heat density distribution; dynamic features, IT load change rate, temperature gradient change; anomaly handling: anomaly data cleaning based on Isolation Forest, etc.

[0065] The spatiotemporal thermal coupling matrix is:

[0066]

[0067] in, Let be the heat load intensity index of cabinet (i,j) at time t. α is the server power consumption weighting coefficient, with an initial value of 0.7. Let be the total power consumption of the servers in rack (i,j) at time t, and β be the thermal gradient weighting coefficient, with an initial value of 0.3. Let t be the temperature difference between the front and back ends of the cabinet (i,j).

[0068] Its construction methods include:

[0069] S21. Define the connection relationship between adjacent cabinets based on the data center topology graph G=(V,E), where V is a cabinet node and E is an adjacent edge;

[0070] S22. The total power consumption of the servers in the rack is combined with the temperature difference between the front and back ends of the rack by weighted summation. The server power consumption weighting coefficient and the thermal gradient weighting coefficient are optimized by backpropagation.

[0071] S3. Establish a spatiotemporal graph convolutional network-long short-term memory (STGCN-LSTM) hybrid prediction model; specifically including:

[0072] S31. Spatial features are extracted using a spatiotemporal graph convolutional layer, with the following formula:

[0073]

[0074] in, To add a self-loop adjacency matrix, Let X be a degree matrix. (l) Let W be the input feature matrix of the l-th layer. (l) Let σ be the trainable weight matrix of the l-th layer, and σ be the ReLU activation function.

[0075] S32. Extract temporal features using an adaptive forgetting gate LSTM layer, the formula is as follows:

[0076] f t =σ(W f ·[h t-1 ,x t ]+b f +C f ·ΔT env );

[0077] Among them, f t W is the output value of the forget gate at time t. f h is the forget gate weight matrix. t-1 Let x be the hidden state vector at time t-1. t Let b be the input feature at time t. f For the forget gate bias term, C f Let ΔT be the environmental disturbance sensitivity coefficient. env The outdoor temperature difference disturbance factor;

[0078] Among them, the outdoor temperature difference disturbance factor ΔT env The calculation method is as follows:

[0079]

[0080] S33. Based on the thermodynamic equation, the physical constraint correction layer is formulated as follows:

[0081]

[0082] in, The corrected load prediction value at time t. Let ∑P be the original load forecast value at time t.IT Let P be the total power consumption of IT equipment at time t, η be the COP coefficient of the air conditioning system, and P be the total power consumption of IT equipment at time t. cool_max The maximum theoretical cooling capacity of the refrigeration system is given; the COP coefficient η of the air conditioning system is obtained by fitting historical data, with the following constraints:

[0083] |∑P IT -η·P cool |≤ε, where ε is the allowable error threshold for energy conservation;

[0084] S4. Use a Seq2Seq structure to perform multi-step rolling prediction and generate prediction results; the formula for multi-step rolling prediction is:

[0085] [Y t+1 ,...,Y t+k ] = Decoder(Encoder(H t-n:t ),C phys );

[0086] Among them, Y t+1 ,...,Y t+k Given a future k-step load prediction sequence, Decoder is a decoder network, Encoder is an encoder network, and H... t-n:t For the historical spatiotemporal thermal coupling matrix sequence, C phys This is the physical constraint vector that includes the pipeline resistance curve;

[0087] Multi-step rolling prediction uses the TeacherForcing mechanism to train a Seq2Seq model;

[0088] S5. Generate air conditioning control commands based on the prediction results, dynamically adjusting the compressor frequency and valve opening. The objective function of the control commands is:

[0089]

[0090] Among them, T k Let k be the predicted temperature, ω1 be the temperature setpoint, ω1 be the temperature stability weight (default value is 0.6), and Δf be the temperature setpoint. k ω is the compressor frequency change at time k, ω2 is the equipment action penalty weight with a default value of 0.4, and N is the optimization time domain length.

[0091] Example 2:

[0092] This embodiment also provides a data center air conditioning equipment load prediction system, including:

[0093] The data acquisition module, configured with multi-source sensors and data interfaces, is used to collect historical operational data from the data center. The module includes: temperature sensors, humidity sensors, air conditioning return / supply water temperature sensors, IT equipment power consumption meters, chiller power meters, outdoor weather stations, etc. Specifically:

[0094] Temperature sensors are deployed at cabinet entrances and exits, cold aisles, and air conditioning water pipes.

[0095] Power consumption monitoring equipment, which connects to IT equipment and chiller units;

[0096] The IoT gateway integrates BMS system data and weather station API, with a sampling frequency of 1 minute / time for environmental data and 5 seconds / time for device power consumption.

[0097] The data processing module is used to perform spatiotemporal alignment on the collected historical operational data, construct a spatiotemporal thermal coupling matrix, and use the isolated forest algorithm to clean up abnormal data.

[0098] Model building module: used to integrate the STGCN-LSTM hybrid model and the physical constraint correction layer to build a hybrid prediction model;

[0099] Prediction module: Used to perform multi-step rolling prediction using a Seq2Seq structure and generate prediction results;

[0100] Control module: Used to generate air conditioning control commands based on prediction results, and dynamically adjust compressor frequency and valve opening;

[0101] The digital twin module is used to generate 3D heat maps to render the temperature distribution in the computer room in real time, and to control the robustness of the strategy in simulation tests of load change scenarios.

[0102] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0103] It should be noted that the components mentioned in the above embodiments are all general standard parts or components known to those skilled in the art. Their structures and principles can be learned by those skilled in the art through technical manuals or conventional experimental methods.

[0104] This invention has illustrated its principles and implementation methods using specific examples. The descriptions of these embodiments are merely illustrative of the method and its core ideas; furthermore, those skilled in the art will recognize that modifications may be made to the specific implementation methods and application scope based on the principles of this invention. Therefore, the content of this specification should not be construed as limiting the invention.

Claims

1. A method for predicting the load of air conditioning equipment in a data center, characterized in that: Includes the following steps: S1. Collect historical operational data from the data center; S2. Perform spatiotemporal alignment on the collected historical operation data, construct a spatiotemporal thermal coupling matrix, and use the isolated forest algorithm to clean up abnormal data; S3. Establish a hybrid prediction model; S4. Use a Seq2Seq structure to perform multi-step rolling prediction and generate prediction results; S5. Generate air conditioning control commands based on the prediction results, and dynamically adjust the compressor frequency and valve opening.

2. The data center air conditioning equipment load prediction method according to claim 1, characterized in that: In step S1, the historical operating data includes server power consumption P. it Temperature difference between cabinet inlet and outlet chiller unit power P cool Outdoor temperature and humidity T out Air conditioning water supply temperature T supply Air conditioning return water temperature T return .

3. The data center air conditioning equipment load prediction method according to claim 1, characterized in that: In step S2, the spatiotemporal thermal coupling matrix is: in, Let be the heat load intensity index of cabinet (i,j) at time t. α is the server power consumption weighting coefficient, with an initial value of 0.

7. Let be the total power consumption of the servers in rack (i,j) at time t, and β be the thermal gradient weighting coefficient, with an initial value of 0.

3. Let t be the temperature difference between the front and back ends of the cabinet (i,j). Its construction methods include: S21. Define the connection relationship between adjacent cabinets based on the data center topology graph G=(V,E), where V is a cabinet node and E is an adjacent edge; S22. The total power consumption of the servers in the rack is combined with the temperature difference between the front and back ends of the rack by weighted summation. The server power consumption weight coefficient and the thermal gradient weight coefficient are optimized by backpropagation.

4. The data center air conditioning equipment load prediction method according to claim 1, characterized in that: Step S3 includes: S31. Spatial features are extracted using a spatiotemporal graph convolutional layer, with the following formula: in, To add a self-loop adjacency matrix, Let X be a degree matrix. (l) Let W be the input feature matrix of the l-th layer. (l) Let σ be the trainable weight matrix of the l-th layer, and σ be the ReLU activation function. S32. Extract temporal features using an adaptive forgetting gate LSTM layer, the formula is as follows: f t =σ(W f ·[h t-1 ,x t ]+b f +C f ·ΔT env ); Among them, f t W is the output value of the forget gate at time t. f h is the forget gate weight matrix. t-1 Let x be the hidden state vector at time t-1. t Let b be the input feature at time t. f For the forget gate bias term, C f Let ΔT be the environmental disturbance sensitivity coefficient. env The outdoor temperature difference disturbance factor; S33. Physical constraint correction layer, the formula is: in, The corrected load prediction value at time t. Let ∑P be the original load forecast value at time t. IT Let P be the total power consumption of IT equipment at time t, η be the COP coefficient of the air conditioning system, and P be the total power consumption of IT equipment at time t. cool_max This represents the maximum theoretical cooling capacity of the refrigeration system.

5. The data center air conditioning equipment load prediction method according to claim 4, characterized in that: In step S32, the outdoor temperature difference disturbance factor ΔT env The calculation method is as follows:

6. The data center air conditioning equipment load prediction method according to claim 4, characterized in that: In step S33, the COP coefficient η of the air conditioning system is obtained by fitting historical data, with the following constraints: |∑P IT -η·P cool |≤ε, where ε is the allowable error threshold for energy conservation.

7. The data center air conditioning equipment load prediction method according to claim 1, characterized in that: In step S4, the formula for the multi-step rolling prediction is: [Y t+1 ,...,Y t+k ]=Decoder(Encoder(H t-n:t ),C phys ); Among them, Y t+1 ,...,Y t+k Given a future k-step load prediction sequence, Decoder is a decoder network, Encoder is an encoder network, and H... t-n:t For the historical spatiotemporal thermal coupling matrix sequence, C phys This is the physical constraint vector that includes the pipeline resistance curve; The multi-step rolling prediction uses the TeacherForcing mechanism to train the Seq2Seq model.

8. The data center air conditioning equipment load prediction method according to claim 1, characterized in that: In step S5, the objective function of the control pointer is: Among them, T k Let k be the predicted temperature, ω1 be the temperature setpoint, ω1 be the temperature stability weight (default value is 0.6), and Δf be the temperature setpoint. k ω is the compressor frequency change at time k, ω2 is the equipment action penalty weight with a default value of 0.4, and N is the optimization time domain length.

9. A data center air conditioning equipment load prediction system, characterized in that: include: The data acquisition module is equipped with multi-source sensors and data interfaces to collect historical operational data from the data center. The data processing module is used to perform spatiotemporal alignment on the collected historical operational data, construct a spatiotemporal thermal coupling matrix, and use the isolated forest algorithm to clean up abnormal data. Model building module: used to integrate the STGCN-LSTM hybrid model and the physical constraint correction layer to build a hybrid prediction model; Prediction module: Used to perform multi-step rolling prediction using a Seq2Seq structure and generate prediction results; Control module: Used to generate air conditioning control commands based on prediction results, and dynamically adjust compressor frequency and valve opening; The digital twin module is used to generate 3D heat maps to render the temperature distribution in the computer room in real time, and to control the robustness of the strategy in simulation tests of load change scenarios.

10. The data center air conditioning equipment load prediction system according to claim 1, characterized in that: The data acquisition module includes: Temperature sensors are deployed at cabinet entrances and exits, cold aisles, and air conditioning water pipes; A power consumption monitoring device, which is connected to IT equipment and a chiller unit; The Internet of Things (IoT) gateway integrates BMS system data and weather station API, with a sampling frequency of 1 minute / time for environmental data and 5 seconds / time for device power consumption.

Images